Authorship Determination Using Letter Pair Frequency Features with Neural Network Classifiers
Bradley P. Kjell · Literary and Linguistic Computing · 1995
The relative frequencies of letter pairs within text samples can be used in authorship studies. Neural network classifiers can easily accommodate the large number of these features that might be used in describing a text. This is tested by using neural networks to recognize the authorship of individual papers from the Federalist Papers. For a wide selection of features and configurations, trained networks had low network errors and no classification errors with papers from the training set. Many networks also showed no errors in classifying papers from an independent testing set. However, classification of the twelve papers of uncertain authorship was inconsistent, even for those networks showing no classification error with the other papers. Since the method is fast and easy, it may be useful when a great amount of text must be screened without taking the time for more painstaking effort.